Where to Start With AI: Rank the Work, Not the Tools

You’ve sat through three demos this month. One tool writes emails, one answers the phone, one promises to run your social media. You signed up for one of them. Six weeks later, two people have logged in, and the monthly charge keeps showing up on the card.

Nothing was wrong with the tool. The order was wrong. You picked the answer before you’d written down the question.

The question is simple, and it’s about your own week. Which work eats the most hours, happens the most often, follows the same steps every time, and won’t hurt anybody if it goes slightly wrong once? Rank that, and the tool shopping gets much shorter.

I’ll tell you what I bring to this, and what I don’t. I haven’t worked your counter, your truck or your front desk. For almost 18 years I ran a digital agency called Infographic World, and these days I use AI to fix real problems inside businesses. That has meant a lot of time watching where the hours go in an office, and they rarely go where the owner thinks. The tasks that pile up are small and boring, and nobody counts them.

Where the hours actually go

Ask an owner where the week goes and you’ll hear about the big things: the difficult customer, the job that went sideways. Those are real, and they’re mostly judgment, the part AI is worst at taking over.

The hours that pile up look different. They’re repeated many times a week, in small pieces, often by someone being interrupted.

• Answering the same handful of questions by phone and text. Are you open, is it ready, how much does it usually cost, do you take this insurance.
• Turning notes into documents. Job notes into an estimate, an estimate into an invoice, a conversation into a follow-up email.
• Copying information from one place to another. A form into the scheduling system, a receipt into the books, an order into a spreadsheet.
• Chasing. Unpaid invoices, unsigned estimates, missing paperwork, unanswered reviews.
• Sorting. The inbox, the voicemail, the stack of receipts, the photos from the job site.

Each one feels like five minutes. Nobody adds them up.

A four-part score

Here’s a scoring method you can run on paper. Take each task and give it a score from 1 to 5 on four things.

1. Hours per week. 1 for under an hour, 2 for one to two hours, 3 for three to four, 4 for five to seven, 5 for eight or more. Add up everyone who does it.
2. Frequency. 1 for monthly or less, 2 for weekly, 3 for a few times a week, 4 for daily, 5 for many times a day. Frequent tasks give you more chances to see whether a change is working.
3. How rule-based it is. 5 if you could write the steps on an index card and a new hire would follow them. 1 if every case takes judgment from your most experienced person.
4. Mistake safety. 5 if a mistake is cheap, easy to spot and easy to fix. 1 if a mistake costs real money, loses a customer, or creates a legal, safety or privacy problem.

Add them up for a total out of 20. Then apply one gate: any task scoring 1 on mistake safety is off the list for a first project, no matter how high the total. Not forever. Just until you’ve done something safer and learned how the tools behave in your business.

A worked example, with made-up tasks

Picture a made-up auto repair shop with nine people. Here’s how five of its tasks might score. Every number below is invented for the example.

• ”Is my car ready?” calls and texts. About 6 hours a week across the front desk (4), many times a day (5), mostly the same answer pulled from the shop system (4), and a wrong status update is annoying but quickly corrected (4). Total: 17.
• Writing estimates from the tech’s notes. About 5 hours a week (4), daily (4), part rules and part judgment (3), and a pricing mistake costs money or trust (2). Total: 13.
• Categorizing expenses for the bookkeeper. About 3 hours a week (3), weekly (2), very rule-based once the categories are set (5), and errors get caught at month end (3). Total: 13.
• Social media posts. About 1.5 hours a week (2), weekly (2), mostly creative (2), and a weak post costs little (5). Total: 11.
• Diagnosing a tricky electrical fault. About 8 hours a week (5), daily (4), pure judgment (1), and a wrong call costs a comeback or worse (1). Total: 11, and gated out.

The status calls win, and not by a little. They’re frequent, they follow rules, and the cost of a mistake is low. That makes them a good place to learn. An assistant that reads the shop system and answers status questions by text, with anything about price or repairs handed to a service writer, is a narrow, testable first project.

Notice what didn’t win. Diagnosis eats the most hours, and it’s the most interesting problem in the building. It’s also the work that most needs your best tech. The estimates and the bookkeeping tie for second. The estimates take more hours, so they’d be the next project once the shop trusts the setup, with a person approving every price.

Run the numbers

The score tells you where to look. This tells you how big the problem is today. It isn’t a savings estimate, and I’d be wary of anyone who hands you one before a pilot.

The formula, in words: hours per week on the task, times the loaded hourly cost of the person doing it, times the weeks they work in a year. That’s what the task costs you now. If the task also produces mistakes, add the number of mistakes per month times the cost to fix each one, times twelve.

Say, for the sake of the math, that the status calls take 6 hours a week, the front desk person costs $25 an hour with taxes and benefits, and they work 50 weeks a year. All made up. That’s 6 × $25 × 50, or $7,500 a year spent answering one question. Run the same math on the estimates with a made-up $30 an hour and you get 5 × $30 × 50, also $7,500, plus whatever the pricing mistakes cost.

Two tasks can cost the same and still deserve different places in line. That’s why the score matters.

The three numbers to pull from your own business:

1. Hours per week on each task. Keep rough notes for one normal week. Tally marks are fine.
2. The loaded hourly cost of the person who does it. Wage plus payroll taxes and benefits. If it’s you, use what your hour is worth when it goes to sales or customers.
3. Mistakes per month on each task, and what one costs to fix. A redo, a refund, a lost customer, an hour of someone’s time.

What has to be in place first

• A week of honest time notes. Not perfect tracking. Rough notes from the people doing the work.
• The data the task touches, in a system. Status calls only work if job status lives in the shop system and not in someone’s head.
• The rules, written down. If you can’t write the steps, the task isn’t as rule-based as you scored it.
• One owner for the project. A named person who checks it weekly and can say stop.

Guardrails worth keeping

• A person approves anything that moves money or makes a promise to a customer.
• Respect the mistake gate. Start where errors are cheap.
• Use business-grade tools. Read the data settings before customer information goes in.
• Keep the baseline. Write down this week’s numbers before you change anything, or you’ll be grading the project on feel.

How you’ll know it’s working

• Hours per week on the chosen task, against the baseline week.
• Mistakes or redos per month on that task.
• Response time, if the task faces customers.
• How many of your people are still using it after 30 days.
• Customer complaints about the task, which should hold steady or drop.

Where I’d start

Spend one normal week keeping rough notes on your five most repetitive tasks. Score them on the four parts, apply the mistake gate, and pick the top scorer that passes. Run it as a 30-day pilot with one person in charge and the baseline written down before day one.

If you want a quick way to rough out the numbers before that week starts, try the AI Opportunity Calculator.